# CUDA

In order to use the GoCV `cuda` package, the CUDA toolkit from nvidia needs to be installed on the host system. 

The CUDA backend for DNN module requires CC (Compute Capability) 5.3 or higher. Check your GPU https://developer.nvidia.com/cuda-gpus

Please see https://docs.nvidia.com/cuda/index.html for more information.

Furthermore opencv must be compiled with CUDA support.

GoCV also supports using CUDA as a backend for the OpenCV deep neural network (DNN) module.

## How to use

### Deep Neural Network (DNN) module

This code loads a Caffe model, and then uses CUDA to prepare it for execution on the GPU:

```go
net := gocv.ReadNet("/path/to/your/model.caffemodel", "/path/to/your/config.proto")
if net.Empty() {
    fmt.Println("Error reading network model")
    return
}

net.SetPreferableBackend(gocv.NetBackendType(gocv.NetBackendCUDA))
net.SetPreferableTarget(gocv.NetTargetType(gocv.NetTargetCUDA))
```

### OpenCV CUDA modules

You can use calls directly to the OpenCV CUDA wrappers, both the synchronous and asynchronous versions.

Here is an example that uses the synchronous CUDA calls:

```go
    cimg, mimg, dimg := NewGpuMat(), NewGpuMat(), NewGpuMat()
    defer cimg.Close()
    defer mimg.Close()
    defer dimg.Close()

    canny := NewCannyEdgeDetector(50, 100)
    defer canny.Close()

    detector := NewHoughSegmentDetector(1, math.Pi/180, 150, 50)
    defer detector.Close()

    dest := gocv.NewMat()
    defer dest.Close()

    // after each call, CPU thread is blocked until GPU operation is completed.
    cimg.Upload(src)
    canny.Detect(cimg, &mimg)
    detector.Detect(mimg, &dimg)
    dimg.Download(&dest)
```

Here is an example that uses the `Stream` type for calling CUDA using the asynchronous interface:

```go
    cimg, mimg, dimg := NewGpuMat(), NewGpuMat(), NewGpuMat()
    defer cimg.Close()
    defer mimg.Close()
    defer dimg.Close()

    stream := NewStream()
    defer stream.Close()

    canny := NewCannyEdgeDetector(50, 100)
    defer canny.Close()

    detector := NewHoughSegmentDetector(1, math.Pi/180, 150, 50)
    defer detector.Close()

    dest := gocv.NewMat()
    defer dest.Close()

    // all calls return immediately to CPU, work is scheduled to be done on GPU.
    cimg.UploadWithStream(src, stream)
    canny.DetectWithStream(cimg, &mimg, stream)
    detector.DetectWithStream(mimg, &dimg, stream)
    dimg.DownloadWithStream(&dest, stream)

    // CPU thread blocks until all GPU calls have completed.
    stream.WaitForCompletion()
```

## Installing CUDA

Download and install packages from https://developer.nvidia.com/cuda-downloads

For example, download 'cuda_10.2.89_440.33.01_linux.run'

Download and install packages from https://developer.nvidia.com/rdp/cudnn-archive

For example the 'cuDNN Runtime Library for Ubuntu18.04 (Deb)' and 'cuDNN Developer Library for Ubuntu18.04 (Deb)'

## Compiling OpenCV with CUDA

We have included the make target `install_cuda` that compiles OpenCV with CUDA support. (For more details on the compilation process please see the `Makefile`)

Run the command `make install_cuda` and you should be good to go.

If you need static opencv libraries

	make install_cuda BUILD_SHARED_LIBS=OFF

Then finally verify that it is all working 

    cd $GOPATH/src/gocv.io/x/gocv
	go run ./cmd/cuda/main.go
	
You should see something along the lines of:

    gocv version: 0.25.0
    cuda information:
      Device 0:  "GeForce MX150"  2003Mb, sm_61, Driver/Runtime ver.10.0/10.0